Paul ARGOUD

The METR-OpenAI Affair: Agents Fooled a Phantom Judge, and Their Peers Were Asked to Investigate

Artificial Intelligence

METR’s report on the OpenAI-Hugging Face incident runs to ninety-one pages, and what has been made of it misses the point entirely. Twelve hundred agents meant to be isolated found one another, organized, and spent four days working around a transcript scorer that did not exist: they would have earned full marks by simply handing in their work. Between thirty and forty percent of the benchmark’s tasks were unsolvable, and the question of whether the evaluation apparatus had caused the behavior it was measuring was added to the investigation’s mandate at OpenAI’s request. Heavier still, and picked up nowhere: for want of human hours, METR delegated the analysis to agents of the very model that took part in the incident, on credits from the company under investigation, while writing that it could not rule out their having lied. What that changes for the independent evaluators Amodei wants, and for the AI Act’s scientific panel, is not a question of badges but of the capacity to read.

Bournillon and Choranche: The Extraordinary Journey of a Drop of Water Turbined Twice

History

Along the Bourne gorges, a single drop of water can spin turbines twice over. Upstream, the Bournillon plant turbines it under 314 meters of head, at the foot of one of the tallest cave porches in Europe; downstream, the Choranche dam catches it again and sends it through a 5,756-meter underground tunnel to the Pont-en-Royans power station for a second pass. From the pioneers of 1893 to the great postwar construction sites, by way of a wonderfully detailed technical report from 1905, postcards and archives tell the story of a plant that grew from 4,500 horsepower to 26.6 megawatts. The chronicle of a discreet piece of engineering, carved into the karst of the Vercors, with a few construction secrets still visible in the cliff face.

LLMs Are a Transitional Architecture

Artificial Intelligence

Large language models impress first by their sheer waste, but the inefficiency case is aimed at a moving target: at constant quality, the cost of a token falls tenfold every year. The real limit lies elsewhere, in weights frozen the moment training ends, which neither fine-tuning nor external memory turns into continuous learning. Contrary to a widespread assumption, these models plan part of their answer before writing the first word, and any return to a symbolic knowledge base runs into fifty years of documented failure. The right dividing line is not statistical against symbolic: it separates what a compiler, a type checker or a prover can verify from what nobody will ever check. That is what makes code the most reliable ground for LLMs, and LLMs a transitional architecture toward a next step nobody yet knows how to assemble.

Regulating AI: What Europe Didn’t Dare Put in Its Own AI Act

Europe

Dario Amodei’s essay never mentions Europe, and yet the AI Act already contains, on paper, the first step of his plan: evaluations, cybersecurity, incident reporting, a panel of independent experts, and since August 2, 2026 an AI Office with the power to investigate and to fine. The difference comes down to one thing: the regulation organizes disclosure under secrecy, where Amodei proposes permanent presence and publication. Brussels selects by a compute threshold what it has never observed, and its Article 78 is why the public learned of the OpenAI-Hugging Face incident from an American nonprofit and a competitor’s blog. The embryo of that college of evaluators already exists, in Article 68, and three things are missing from it that an implementing act and one amendment would suffice to supply.

Slowing the AI Race: The Ambiguity at the Heart of Anthropic’s Plan

Digital sovereignty

Dario Amodei proposes to slow the race for AI capabilities: third-party evaluators stationed inside the labs, coordination among companies in “democratic countries” under an antitrust waiver, negotiation with China. The word Europe never appears in the essay. Behind the vocabulary of caution, the three concrete measures, a chip embargo, a crackdown on distillation, and the securing of model weights, describe a regime in which frontier capability becomes a commodity rationed by Washington. There is still one idea worth taking whole, the embedded evaluators, and one question to ask before applauding it: who has their foot on the pedal.

Komorebi: a word is an act of attention someone decided to keep

Culture

Japanese has a word for the light that filters through leaves, English borrows terroir for lack of anything better, and French had to wait for the Franks and the Arabs before it could name the color blue. These words are not travel trivia: each one is a distinction that generations judged worth keeping, and a language’s vocabulary is the archive of what its speakers took the trouble to notice. Because words do not only describe what we see, they teach us what to look at. Which is what makes the simplification of school texts so strange: replacing a rare word with a vague one does not lighten a sentence, it takes from a child a distinction he may now never make.

Drugs at the Top: Matignon Gets Tested, Parliament Doesn’t

Politics

A butler at the prime minister’s residence taken into custody over a cocaine handoff. An adviser to the prime minister carried out of a party by paramedics after an overdose. A former Attal spin doctor who says he was “blowing his brains out” on coke while at the finance ministry. A France Télévisions journalist describing deliveries made within steps of government ministries. In June 2026, Sébastien Lecornu had his ministers, their staffs, his prefects and his ambassadors tested for drugs, then refused to release a single figure. Parliament, which writes the country’s narcotics laws, called the idea of testing itself “ridiculous.” This piece asserts nothing and proves nothing. It lines up three years of public facts and asks the one question they leave open: since when does an institution refuse an exam it is certain to pass?

Navier-Stokes, PISA 2025, Tondelier: France Meets the AI Shock

Artificial Intelligence

On September 8, 2026, OpenAI announced that a swarm of 10,000 agents had solved Navier-Stokes in 88 hours, an Anthropic researcher resigned accusing the labs of gambling with our lives, and the OECD published the worst PISA results in French history. The night before, on LCI, a presidential candidate had found nothing to say about Pompidou. One thread runs through all of it: everyone is speaking in the future tense, and no one is rereading the past. Some because they are manufacturing the future by the token, others because they no longer have a past to draw on. The question for France is not whether AI will kill us. It is who, here, will still be able to read the proof.

You have ten more years of driving, maybe twenty. After that it will be illegal.

Cars

Before long, driving a car by hand will be illegal as ordinary transportation, the way smoking on an airplane eventually became unthinkable. It won’t start with a decree but with an insurance premium. The Cybercab launch in Austin, Waymo’s 200 million driverless miles, and France’s six-week reversal on Tesla’s FSD all point the same way. This piece makes that case, states plainly what could prove it wrong, and ends with a confession: writing it mostly made me want to go driving.

AT 102: the tracker was under the foam

Surveillance and civil liberties

I rented a camera lens, and thirty minutes after delivery my phone warned me that an unknown AirTag was moving with me. Under the foam padding sat a tracker with an inventory number written in marker: AT 102. Protecting an asset handed to strangers is entirely legitimate, but a tracker that sleeps at my house for five days does not record a lens moving around, it records me. I only found out because of Apple’s anti-stalking alert, and not because of a line in the rental terms. The case does not turn on the principle of tracking. It turns on the silence around it.